How to make AI replies not sound like AI

The patterns that give AI-written replies away (dashes, triplets, stock phrases, invented numbers) and the rules ReplyKit's prompt uses to avoid them.

People spot it in one line

Anyone who reads replies on X or comments on LinkedIn has learned the shape of machine-written text. They may not be able to say why, but they scroll past it, and the author does not answer. A reply that reads as AI is worse than no reply: it tells the people you want to impress that you did not bother.

The good news is that the giveaways are a short, fairly stable list. Most of them are habits of style, which a prompt can forbid and a human can fix in ten seconds. This guide is the list ReplyKit's own prompt is built on.

Punctuation and rhythm

Em dashes. The long dash is the single most recognised tell. Use a period or a comma instead.
Even sentences. Models write sentences of similar length. People do not. Put a three-word sentence next to a long one.
Lists of three. "Fast, cheap and reliable." Models reach for triplets by reflex. Use one item, two, or four.
A trailing -ing clause. "…, highlighting the importance of trust." It adds the look of depth and no meaning. Cut it.

Sentence shapes

Negative parallelism. "It's not about X, it's about Y." "Not just A but B." Say Y directly.
The clipped closer. A fragment tacked on at the end for punch: "No fluff." "Simple as that."
Inflated verbs. "Serves as", "stands as", "represents", "boasts". Use is, are and has.
The upbeat ending. "Exciting times ahead!" "The future is bright." End when the point is made.

Words and phrases that give it away

Some words appear in machine text far more often than in human writing: delve, tapestry, testament, pivotal, crucial, vibrant, foster, showcase, elevate, unlock, harness, seamless, robust, game-changer, and "landscape" or "navigate" used figuratively.

Some phrases are worse, because they pretend to insight: "the real question is", "the thing most people miss", "nobody talks about this", "X is just Y with extra steps", "moves the needle", "speaks volumes", "it's worth noting", "at the end of the day". A reader who has seen them a hundred times stops reading at the first one.

Content tells, which matter more

Style is the easy part to fix. The deeper tells are about what the reply says.

Generic agreement. "Great point!" followed by the post restated in other words. It adds nothing, so it reads as filler.
Invented specifics. "Studies show 73% of founders…" A made-up number is worse than none, and people check.
The lecture. Explaining the author's own topic back to them, usually with a numbered list.
Abstract over concrete. "Many people these days" where a person would name the tool, the client or the day it happened.

What reads as human

The opposite habits are simple to state. Pick one specific thing in the post and respond to that. Bring a detail only you would know. Keep it short. Allow a little imperfection: a fragment, an aside in brackets, a sentence that starts with "and" or "but". Write in the post's language and register; a casual lowercase post deserves a casual reply.

And only state facts you are sure of. Details taken from the post itself are always safe.

A 10-second edit before you post

Whatever tool wrote the draft, read it once with this list:

Does the first line respond to something specific in the post?
Any dash, triplet, or phrase from the lists above? Rewrite that bit.
Is every number and name true?
Would you say this out loud to the author? If not, change the words that are not yours.

How ReplyKit handles it

Every rule above is written into ReplyKit's prompt, for X, LinkedIn and Reddit, on every plan including free. Each click gives three drafts with different angles, so you pick the one closest to what you think instead of accepting the only option. With voice learning on, it reads your own recent X posts and writes the way they read.

You can test it on any post with the free generators, no account needed.

ReplyKit is a solo project. Tell me what you need.

There is no company behind this, just one developer who reads every message. If ReplyKit does not fit how you work, a feature, a platform, a different style of reply, say so. Requests from users decide what ships next.

Prefer email? contact@replykit.eu